The core of the work is an enhanced convolutional autoencoder, a neural network architecture that has become one of the workhorses of modern engineering simulation. An autoencoder consists of two paired networks: an encoder that squeezes a high-dimensional input, in this case a detailed axial temperature profile of the igniter combustion chamber, into a much smaller latent representation, and a decoder that attempts to rebuild the original field from that compressed code. When the two networks are trained together on large sets of simulated or measured temperature fields, the latent space learns to retain only the information that genuinely matters for describing the physics, discarding redundancy. The result is a compact mathematical fingerprint of each thermal state that can be stored, transmitted or fed into downstream models at a tiny fraction of the original size.

Dimensionality reduction is not a new idea in combustion science. For decades engineers have relied on proper orthogonal decomposition, or POD, a linear technique that decomposes a set of flow or temperature fields into orthogonal modes ranked by their energy content. Truncating the mode series yields a reduced-order model that captures the dominant behavior with far fewer degrees of freedom than the full simulation. POD has proven remarkably effective for many fluid and thermal problems, but it has a fundamental limitation: it is linear. Combustion is anything but. The coupling between chemical reactions, turbulent mixing and heat release produces strongly nonlinear relationships between operating conditions and the resulting temperature distribution, and linear mode superpositions struggle to represent that structure efficiently at high compression ratios.

This is precisely where the convolutional autoencoder earns its keep. Because the encoder stacks nonlinear convolutional layers with activation functions, it can warp and fold the data manifold in ways that no linear decomposition can match. The Chinese team enhanced the basic architecture to balance two competing objectives: reconstruction accuracy and the dimensionality of the latent code. Push the compression too far and fine thermal features blur away; keep too many latent variables and the model offers little advantage over the raw field. By tuning the architecture, the researchers achieved a compression ratio of up to 32 to 1, meaning that a full axial temperature field can be represented by roughly one thirty-second of its original information content while still being reconstructed with engineering-grade fidelity.

The accuracy figures reported in the study are notable in themselves. Under noisy inputs and across a range of oxygen-to-fuel ratio conditions, the relative errors for both the average chamber temperature and the proportion of high-temperature zones remained below six percent. These two quantities are not arbitrary metrics. The average temperature characterizes the overall thermal load on the igniter structure, while the fraction of high-temperature zones indicates where hot spots concentrate, which directly governs material selection, cooling strategy and fatigue life. A reduced-order model that preserves both quantities within a six percent band is therefore not merely compressing data; it is preserving the engineering quantities that matter for design decisions.

Robustness was a central concern of the work. Real sensor data and real simulations are never clean. Measurement noise, numerical artifacts and fluctuations in operating conditions all contaminate the inputs that a surrogate model must digest. The researchers explicitly tested their network under noise interference and under varying mixture ratios between the gaseous oxygen and gaseous methane streams. The model maintained its performance across these perturbations, demonstrating the generalization capability that separates a useful engineering tool from a fragile laboratory demonstration. In the unforgiving context of rocket propulsion, where a failed ignition can destroy an entire mission, that robustness is not a luxury but a requirement.

Perhaps the most intriguing finding is that the compressed latent space spontaneously exhibits a degree of physical interpretability, even though the researchers imposed no explicit physical constraints on the network. In many machine learning applications the latent variables of a deep network are opaque, an inscrutable set of numbers with no obvious meaning. Here, without being told anything about combustion physics, the autoencoder organized its internal representation in ways that correlate with meaningful physical structure. This echoes a growing theme in the reduced-order modeling literature, where physics-assisted and physics-aware autoencoders have been used for turbulence modeling and for reconstructing liquid rocket engine flames. The fact that interpretability can emerge without explicit constraints suggests that the network is discovering genuine structure in the data rather than memorizing examples.

The comparison with proper orthogonal decomposition at the same compression rate is the study’s benchmark moment. At equal compression, the convolutional autoencoder delivered higher reconstruction accuracy and superior generalization performance than the classical linear method. This outcome matters beyond the specific case of igniters. It provides further evidence that nonlinear neural compression is becoming the preferred tool for combustion and thermal problems where the relationship between operating parameters and field variables is strongly nonlinear. Related work in the field, from machine-learning-based reduced-order models of unsteady flows to Wasserstein autoencoder-enhanced models for turbine blade life monitoring, points in the same direction: deep encoders are steadily displacing linear decomposition as the default compression layer in engineering surrogates.

The practical implications extend to how rocket engines will be designed and monitored in the future. High-fidelity combustion simulations of even a small igniter chamber are computationally expensive, often requiring detailed chemical kinetics mechanisms such as GRI-Mech and resolving turbulent flame structure over millions of grid points. Running enough of these simulations to explore the full design space of mixture ratios, geometries and transient startup sequences is prohibitive. A trained autoencoder changes that calculus. Once the encoder-decoder pair has been fitted, each new thermal state can be compressed, compared and reconstructed almost instantly, enabling rapid design iteration, real-time condition monitoring and anomaly detection. Combustion stability monitoring through flame imaging and autoencoder-based networks has already been demonstrated in power boilers, and the present work brings the same capability to liquid rocket engine hardware.

The authors are explicit that this model is a foundation rather than a finished product. The next step is to couple the compressed latent representation with spatiotemporal prediction models, building toward a dynamic temperature-field prediction framework for combustion engines. In such a framework, the autoencoder handles the spatial compression while a sequence model forecasts how the latent variables evolve in time, allowing the full temperature field to be predicted forward without ever reconstructing the high-dimensional state during the prediction step. If that integration succeeds, it would mark a significant step toward digital twins of rocket engine components: lightweight computational replicas that track the thermal state of real hardware in real time. For a discipline in which every launch depends on a few hundred milliseconds of reliable ignition, teaching machines to see the heat inside an igniter, and to remember it in one thirty-second of the space, is a quietly consequential advance.

Subject of Research: Deep learning-based nonlinear dimensionality reduction for reconstructing temperature fields in rocket igniter combustion chambers

Article Title: Nonlinear Dimensionality Reduction for Temperature Field Reconstruction in Rocket Igniters Based on an Enhanced Convolutional Autoencoder

Article References: Zhang, Y., Wu, X., Zhu, W., Zhang, B., Bao, Q., Zhu, X., Zhang, M., & Qi, H. (2026). Nonlinear Dimensionality Reduction for Temperature Field Reconstruction in Rocket Igniters Based on an Enhanced Convolutional Autoencoder. International Journal of Aeronautical and Space Sciences. https://doi.org/10.1007/s42405-026-01276-7

Image Credits: AI Generated

DOI: 10.1007/s42405-026-01276-7

Keywords: convolutional autoencoder, dimensionality reduction, rocket igniter, temperature field reconstruction, reduced-order modeling, proper orthogonal decomposition, deep learning, combustion, rocket engine, latent space, surrogate modeling, aerospace engineering

The challenge the researchers set out to solve is fundamentally one of missing information. In a tunnel equipped with thermocouples, temperature sensors are typically placed at limited positions along the ceiling and walls because instrumenting every meter of a kilometer-long passage is impractical and expensive. Those sparse readings leave vast stretches of the temperature field unknown. Traditional approaches to filling in the gaps fall into two camps: purely data-driven machine learning models, which can interpolate between sensor points but often violate physical laws, and full computational fluid dynamics simulations, which respect physics but are far too slow to run during an actual emergency. The research team’s answer was to build a model that is both fast and physically faithful, learning from few measurements while remaining constrained by the governing equations of heat transport.

The centerpiece of the study is a Multi-source Physics-informed Kolmogorov–Arnold Network, abbreviated PIKAN. The Kolmogorov–Arnold network, or KAN, is a relatively recent alternative to conventional multilayer perceptrons. Instead of stacking layers of fixed activation functions applied to weighted sums, a KAN places learnable activation functions on the edges of the network, parameterized as one-dimensional functions. This design, rooted in the Kolmogorov–Arnold representation theorem, gives the architecture a different and often more efficient way of approximating complex multivariate functions. In this application, the authors strengthened their KAN by incorporating radial basis function layers, which enhance the network’s nonlinear approximation capacity and, crucially, ensure stable automatic differentiation, the mathematical machinery that allows a neural network to compute its own derivatives with respect to its inputs.

Why do derivatives matter so much here? The answer lies in the physics-informed part of the framework. The researchers embedded the tunnel ceiling heat diffusion equation directly into the loss function of the network, forcing the model’s predictions to satisfy the same partial differential equation that governs how heat spreads along the tunnel ceiling during a fire. Because the network must be differentiated to evaluate that equation, stable automatic differentiation is not a luxury but a necessity. Any numerical noise in the derivatives would corrupt the physics residual and destabilize training. By combining radial basis function layers with the KAN structure, the team obtained smooth, differentiable representations of the temperature field that could be checked against the diffusion equation at thousands of collocation points throughout the tunnel, even where no sensor exists.

The architecture itself is a coordinated ensemble rather than a single monolithic network. It comprises one temperature field network and multiple heat release rate networks, each dedicated to a candidate fire source, and all of them are jointly trained. The temperature field network learns the spatiotemporal distribution of temperature rise throughout the tunnel, while the heat release rate networks infer the intensity and location characteristics of each individual burning source. Training proceeds against two kinds of constraints simultaneously: the observed temperatures from sparse thermocouples, which anchor the solution to reality, and the embedded heat diffusion physics, which keeps the solution plausible everywhere else. The result is a joint inversion in which fire source locations, heat release characteristics, and the complete temperature rise field are all estimated together, each informing the others through the shared physical model.

Validation came from two demanding test cases. The first was a full-scale double-source tunnel fire experiment, the kind of physical trial that reproduces realistic fire behavior at true dimensions rather than in miniature. The second was a three-source tunnel fire scenario simulated with Fire Dynamics Simulator, the widely used computational tool for fire-driven fluid flow. These two cases complement each other: the full-scale experiment demonstrates that the method works against messy, real-world data, while the simulated three-source case probes whether the framework scales to more complex multi-source configurations with known ground truth for rigorous error assessment. Few studies in the fire safety literature attempt this kind of multi-source inversion with both experimental and numerical verification at these levels of complexity.

The reported performance is striking. In the double-source experimental case, the localization errors for the two fire sources were reduced to 0.89 meters and 1.06 meters respectively, meaning the algorithm could pinpoint each blaze within roughly a meter of its true position. In the more challenging three-source simulated scenario, all localization errors were controlled within 1.07 meters, with the minimum error reaching just 0.02 meters, effectively exact localization for one of the three sources. Beyond position, the reconstructed temperature fields agreed well with both the experimental measurements and the numerical simulation results, indicating that the model does not merely find the fires but correctly maps the thermal landscape they create. In direct comparison with the conventional Multi-source Physics-Informed Neural Network, or PINN, the PIKAN approach achieved higher localization accuracy and more reliable temperature reconstruction across the board.

The implications for tunnel safety are substantial. Knowing where a fire is burning, and how many fires are burning, determines where to direct ventilation to avoid pushing smoke toward evacuees, which exits remain tenable, and where firefighters should stage their approach. A temperature field reconstruction that is physically consistent, rather than a statistical guess, gives engineers confidence in the intermediate values between sensors, which matter for assessing structural damage to tunnel linings and for predicting when critical temperatures might be reached at specific locations. Because the method works from sparse thermocouple data already available in many modern tunnels, it could potentially be integrated into existing monitoring infrastructure, turning arrays of temperature sensors into a distributed fire intelligence system. The work also builds on a growing body of research by the same group and others applying intelligent algorithms, from support vector machine fusion for tunnel fire warning to BP neural network-based prediction of ceiling temperatures verified in full-scale experiments.

More broadly, the study signals a shift in how physics-informed machine learning is being engineered. Conventional PINNs, built on standard multilayer perceptrons, have sometimes struggled with spectral bias, an tendency to learn low-frequency components of a solution first and converge slowly on sharp or multi-scale features, exactly the kind of features that characterize the steep thermal gradients near a fire source. Kolmogorov–Arnold networks, with their learnable edge activation functions and enhanced approximation capacity through radial basis functions, represent a promising remedy, and this study is among the first to demonstrate their advantage in a real safety-critical inverse problem involving multiple simultaneous sources. As the authors’ results show, a neural network that respects the heat diffusion equation and possesses the right approximation machinery can turn a handful of ceiling temperature readings into a complete, accurate, physically consistent picture of a multi-source tunnel fire, a capability that could one day mean the difference between a managed evacuation and a catastrophe.

Subject of Research: Physics-informed neural networks for fire source localization and temperature field reconstruction in multi-source tunnel fires

Article Title: Physics-informed Kolmogorov–Arnold network for fire source localization and temperature field reconstruction in multi-source tunnel fires

Article References: Li, Y., & Sun, B. (2026). Physics-informed Kolmogorov–Arnold network for fire source localization and temperature field reconstruction in multi-source tunnel fires. Applied Intelligence, 56(15), Article 441. https://doi.org/10.1007/s10489-026-07504-7

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07504-7

Keywords: tunnel fire, fire source localization, temperature field reconstruction, Kolmogorov–Arnold network, physics-informed neural network, PIKAN, heat diffusion equation, inverse problem, Fire Dynamics Simulator, machine learning, fire safety, sparse sensor data